See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/265142104

Extreme Value Theory: An Introduction

Book · January 2006

DOI: 10.1007/0-387-34471-3

CITATIONS READS
2,301 6,625

2 authors, including:

Ana Ferreira
Universidade de Évora
21 PUBLICATIONS   3,057 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Ana Ferreira on 02 August 2020.

The user has requested enhancement of the downloaded file.

 (correction 2nd part made by MFH)
Extreme Value Theory: An Introduction
by Laurens de Haan and Ana Ferreira

With this webpage the authors intend to inform the readers of errors or mistakes found in the book after publication. We also give extensions
for some material in the book. We acknowledge the contribution of many readers.

Chapter 1: Limit Distributions and Domains of Attraction

Chapter 2: Extreme and Intermediate Order Statistics

Chapter 3: Estimation of the Extreme Value Index and Testing

Chapter 4: Extreme Quantile and Tail Estimation

Chapter 6: Basic Theory in higher dimensional space

Chapter 7: Estimation of the Dependence Structure

Chapter 8: Estimation of the Probability of a Failure Set

Chapter 9: Basic Theory in C[0,1]

Chapter 10: Estimation in C[0,1]

Appendix B: Regular Variation and Extensions

Further and Updated References

Data files

1 /14 11/21/2010
 (correction 2nd part made by MFH)
Chapter 1: Limit Distributions and Domains of Attraction

page line error/unclear/missing correction

8 -0 (complement) Corollary 1.1.3A For x>0
lim a (tx) = x γ
t →∞
a (t )
12 14 log log n + log(4π ) log log n + log(4π )
bn' = (2 log n)1/2 - bn' = (2 log n)1/2 -
(2 log n)1/ 2 2(2 log n)1/ 2
20 -3 … estimator of γ (Section 3.5). … estimator of γ (Section 3.5).
The necessary and sufficient condition for a distribution function to belong to the
Next we show…. domain of attraction of an extreme value distribution is sometimes called “the
extreme value condition”.
Next we show….

Chapter 2: Extreme and Intermediate Order Statistics

page line error/unclear/missing correction

40 -2 …, as we shall see. …, as we shall see.
The asymptotic behavior of intermediate order statistics is important for statistics
The following result …
of extreme values since any meaningful estimator is based on (extreme and)
intermediate order statistics (see Chapter 3).
The following result …
42 10 ⎛ n ⎞ ⎛ nU k +1,n ⎞
k ⎜⎜ − 1⎟⎟ k ⎜ − 1⎟
⎝ kU k +1,n ⎠ ⎝ k ⎠

2 /14 11/21/2010
 (correction 2nd part made by MFH)
Chapter 3: Estimation of the Extreme Value Index and Testing

page line error/unclear/missing correction

77 6 λ λ
ρ c2 ρ c2
104 1 Ψ γ −, ρ ( x ) ≤ ε xγ − + ρ Ψ γ −, ρ ' ( x ) ≤ ε xγ − + ρ '

111 7 Then (3.6.5) becomes Then (3.6.5) (multiplied by f(t)) becomes

10 and (3.6.6) becomes and (3.6.6) (multiplied by f(t)) becomes

113 1 The “negative Hill estimator” was The “negative Hill estimator” was proposed by Smith and Weissman (1985).
proposed by Falk (1995).

Chapter 4: Extreme Quantile and Tail Estimation

page line error/unclear/missing correction

128 label vertical log (1+ γ x) / x log (1+ γ x) / γ
axis in
Fig.4.2(b)
130 -7 ... the moment estimator of γ . We ... the moment estimator of γ . In order to find an estimator for the scale a (n/k) we
define… use relation (3.5.3) for j = 1 and define…
134 -12 Theorem 4.3.1 Theorem 4.3.1 (de Haan and Rootzén (1993))
135 -5 U ( tx ) − U ( t ) U ( tx ) − U ( t )
a (t ) a0 ( t )
138 10 Theorem 4.3.8 Theorem 4.3.8 (Dijk and de Haan (1992))
140 8 1 + 4γ + 5γ 2 + 2γ 3 + 2γ 4 1 + 4γ + 5γ 2 + 2γ 3 = (1 + γ ) (1 + 2γ )
2

3 /14 11/21/2010
 (correction 2nd part made by MFH)

Chapter 6: Basic Theory in higher dimensional space

page line error/unclear/missing correction

209 -8 U1 ([ n ]) etc. U1 ( n ) etc.

213 -12 Remark 6.1.8 Relations (6.1.15) does not Relations (6.1.15) can be written
hold for all Borel sets Ax,y . P ( Ac ) = exp −ν ( A
x, y { x, y )} (6.1.15*)
Where P is the probability measure with distribution function G0. Relation
(6.1.15*) does not necessarily hold for Borel sets A not of the form (6.1.16)

-9 … > 0 and ν ( ∂ A) = 0, and any … … > 0, and any …

216 1 to 9 (6.1.24) … (6.1.24)
dr
π /2
⎛ cos θ sin θ ⎞
= ∫∫r>
x
∧
y
r2
Ψ (dθ )
∫ ⎜⎝ ∨ ⎟ Ψ ( dθ ) .
cosθ sin θ
…
0
x y ⎠
π /2
⎛ cos θ sin θ ⎞
= ∫ ⎜⎝ ∨ ⎟ Ψ ( dθ ) .
0
x y ⎠
217 -11 Definition 6.1.13 We call … Definition 6.1.13 A distribution function G is called a max-stable distribution if
there are constants An, Cn >0, Bn and Dn such that for all x,y and n= 1,2,…
Gn (An x + Bn, Cn y + Dn) = G(x,y) .
It is easy to see that any distibution function G satisfying (6,1,25) is max-stable
and also G(αx + β , γ y + δ) where α , γ are arbitrary positive constants and β , δ
arbitrary real constants. Since any max-stable distribution is in the class of limit
distributions for (6.1.1), we get all the max-stable distributions this way. The class
of…
224 -8 Clearly GL( ) and GU( ) … Clearly GL( ) and GU( ) are max-stable distributions with marginal distributions
n n n n

exp ( − c x ) where c is a generic constant and there …

4 /14 11/21/2010
 (correction 2nd part made by MFH)
231 6 converges to 2-1 … converges to g(x,y) : = 2-1 …

8 = - log G0(x,y) with G0 from Theorem

6.1.1.
= ∫∫{ ∪{t > y}
s > x}
g ( s, t )dsdt for x,y ∈ .

232 5 r3 q (θr, r(1- θ)) . r3 q (θr, r(1- θ)) = q (θ, (1- θ)) (cf. Coles and Tawn (1991)).

15 Let {r i,j}di, j=1 be a matrix Let {r i,j}i=1,2; j=1,2,…,d be a matrix

16 the random vector (V d j=1 r1,j Vj, …, V d j=1 the random vector (V d j=1 r1,j Vj, V d j=1 r2,j Vj )
rd,j Vj )

18 two-dimensional simple distribution two-dimensional distribution function with Fréchet marginals can be
function can be

-9 ∨ twice ∧ twice
233 6 complement 6.14. Let (X,Y) have a standard spherically symmetric Cauchy distribution. Show
that the probability distribution of (|X|,|Y|) is in the domain of attraction of an
extreme value distribution with uniform spectral measure Ψ. Show that the
probability distribution of (X,Y) is also in a domain of attraction. Find the limit
distribution.

Chapter 7: Estimation of the Dependence Structure

page line error/unclear/missing correction

252 15 ... a max-stable distribution function. ... a max-stable distribution function. The marginal distribution are
G 0 ( x, ∞) = exp− a1
x ( )
and G 0 (∞, y ) = exp− a2
y ( )
for some positive a1 and a2

not necessarily one. Hence G 0 is not necessarily simple max-stable (cf.

5 /14 11/21/2010
 (correction 2nd part made by MFH)
Definition 6.1.13).
261 1,2 Not all inequalities have to hold, but at least one of them
262 13 Q ( x, ∞ ) , Q ( x, ∞ ) = 0 ,
14 Q ( ∞, y ) . Q ( ∞, y ) = 0 .

263 -12 = →
265 18 … = x + y - L(x,y) . … = x + y - L(x,y) = R (x,y).

-10 … does not imply asymptotic … does not imply asymptotic dependence.
independence.
268 -2 EW (x1, …, xd) W (x1, …, xd) = μ (R (x1, EW (x1, …, xd) W (y1, …, yd) = μ (R (x1, …, xd) ∩ R (y1, …, yd))
…, xd) ∩ R (x1, …, xd))
269 4 and N is a standard normal random and N indicates a normal probability distribution.
variable.

Chapter 8: Estimation of the Probability of a Failure Set

page line error/unclear/missing correction

273 -0 complement i.e. Qn = cn S (assumption), where …
275 2 more detail below; cf. Theorems 8.2.1 and more detail in section 8.2.; cf. (8.2.7), (8,2,8) and (8,2,15))
8.3.1
283 -7 log 2 ( cn x ) log 2 ( cn x )
2 2 k

6 /14 11/21/2010
 (correction 2nd part made by MFH)

Chapter 9: Basic Theory in C[0,1]

page line error/unclear/missing correction

301 -11 > 0 and υ ( ∂A ) = 0 , and any a > 0 , > 0 and any a > 0 ,
304 -1, -2, -3 ζ 1 , ζ 2 , ζ 3 ,... ( Z1 , π 1 ) , ( Z 2 , π 2 ) , ( Z3 , π 3 ) ,...
be a realization of the point process. be a realization of the point process where Z1 ∈ ( 0, ∞ ] and π 1 ∈ C1+ [ 0,1] . Define
Define ∞
∞
η := ∨ ζ 1 . η := ∨ Ziπ i .
i =1
i =1

306 7 to19 These lines should be indented (belong to (2)).

11 First we note that this implies First we note that this assumption implies
307 -12 d k k
η (s) = ∨ ζ i (s) η (s) = ∨ ζ i (s)
i =1 i =1
-6
Corollary 9.4.5 Corollary 9.4.5 (cf. M. Schlather (2002))

All simple max-stable processes in All simple max-stable processes η in C + [ 0,1] can be generated in the following
C [ 0,1] can be generated in the following way.
+

way.
-4
stochastic processes V1, V2,… in C + [ 0,1] stochastic processes V1, V2,… in C + [ 0,1] := { f ∈ C [ 0,1] : f ≥ 0}

308 6 … Resnick (1977). … Resnick (1977). Let W* be two-sided Brownian motion:

⎧ W + ( s ) for s ≥ 0
W * ( s ) := ⎨ −
⎩W (− s ) for s < 0
+ -
where W and W are independent Brownian motions.

In the rest of the example change W into W* .

311 -5 Theorem 9.5.1 Theorem 9.5.1 (de Haan and Lin (2001))

7 /14 11/21/2010
 (correction 2nd part made by MFH)
315 5 Theorem 9.6.1 (Resnick and Roy Theorem 9.6.1 (Resnick and Roy (1991) and de Haan(1984))
(1991))
320 11 … of the theorem is easy. … of the theorem is easy. ■
{( Z , T )} be a realization of a Poisson point process on
∞
Next we turn … Corollary 9.6.8A Let i i
i =1

( 0, ∞ ] × with mean measure ( dr / r ) × d λ ( λ Lebesgue measure ) . If η is a

2

simple max-stable process in C + ( ) , then there exists a family of functions

f s ( t ) ( s, t ∈ ) with
1. for each t ∈ we have a non-negative continuous function
f s ( t ) : → [ 0, ∞ ) ,
2. for each s ∈

∫ f ( t ) dt = 1 ,
−∞
s (9.6.7A)

3. for each compact interval I ∈

∞

∫ sup f ( t ) dt < ∞ ,
−∞ s∈I
s

⎧ ∞
⎫
{η ( s )}s∈ = ⎨∨ Zi f s (Ti )⎬
d
such that . (9.6.8A)
⎩ i =1 ⎭s∈
Conversely every process of the form exhibited at the right-hand side of (9.6.8A)
with the stated conditions, is a simple max-stable process in C + ( ) .
Proof. Let H be a probability distribution function with a density H` that is
positive for all real x. With the functions fs from Theorem 9.6.7 define the
functions f s ( t ) := f s ( H ( t ) ) H ' ( t ) . Since for any s1 , s2 ,..., sd ∈ and
x1 , x2 ,..., xd positive
∞ f si ( t ) 1 f si ( t )
∫ max
−∞ 1≤ i ≤ d xi
dt = ∫ max
0 1≤ i ≤ d xi
dt ,

8 /14 11/21/2010
 (correction 2nd part made by MFH)

the representation of the corollary follows easily from that of Theorem 9.6.8. ■
Next we turn …

14,16,20,24 [0,1]
1
∞
18, 22, 25 ∫
0 ∫
−∞

321 16, 18, 21, 22, 1 ∞

23, 24 ∫
0
∫
−∞

26 distributions. distributions. ■
*
323 -1 (2#), W W
-6, -8

-6 … independent Brownian motions. … independent two-sided Brownian motions (cf. correction to Example 9.4.6) .

324 5, 6 ⎛ u y−x⎞ ⎛ u y−x⎞

e − x Φ ⎜⎜ − ⎟ e − x Φ ⎜⎜ + ⎟
⎝ 2 u ⎟⎠ ⎝ 2 u ⎟⎠
1 (3#) W W*
2 (2#)
3 (3#)
11 (2#)
12 , 13

-5 x y
325 10 (2#) W W*
11 (2#)
12 (2#)

9 /14 11/21/2010
 (correction 2nd part made by MFH)
-6 Hence for s1 < 0 < s2 Hence for s1 < 0 < s2 and in fact for all real s1, s2
326 2 Let W be Brownian motion independent Let W* be two-sided Brownian motion:
of Y. Consider the process… ⎧ W + ( s ) for s ≥ 0
W ( s ) := ⎨ −
*

⎩W (− s ) for s < 0
where W and W are independent Brownian motions. Let Y and W* be
+ -

independent. Consider the process…

In the rest of the example change W into W* .

-14 a>0 a>1

326- Example 9.8.2 Remove the text of the example Example 9.8.2 (extension of Brown and Resnick (1977)) Let { X ( s )}s∈ be an
328
Ornstein- Uhlenbeck process, i.e.,

( )
X ( s ) = 1{s ≥0} e − s / 2 N + ∫ eu / 2 dW + ( u )
s

( (u ))
−s
+ 1{s <0} e s / 2 N + ∫ eu / 2 dW −
0

with N, W and W independent , N a standard normal random variable and W+

+ -

and W - standard Brownian motions. Since for s ≠ t the random vector (X(s),
X(t)) is multivariate normal with correlation coefficient less than 1, Example 6.2.6
tells us that relation (9.5.1) can not hold for any max-stable process in C[0,1]:
since Y has continuous sample paths, Y(s) and Y(t) can not be independent. Hence
we compress space in order to create more dependence, i.e., we consider the
convergence of
⎧⎪ n ⎛ ⎛ s ⎞ ⎞ ⎫⎪
⎨∨ bn ⎜⎜ X i ⎜ 2 ⎟ − bn ⎟⎟ ⎬ (9.8.4)
⎩⎪ i =1 ⎝ ⎝ bn ⎠ ⎠ ⎭⎪s∈
in C [ − s0 , s0 ] for arbitrary s0 > 0, where X1, X2, … are independent and identically
distributed copies of X and the bn are the appropriate normalizing constants for
the standard one-dimensional normal distribution, e.g., (cf. Example 1.1.7)
bn = ( 2 log n − log log n − log ( 4π ) ) . Then
1/ 2

10 /14 11/21/2010
 (correction 2nd part made by MFH)
⎛ ⎛ s ⎞ ⎞
bn ⎜⎜ X ⎜ 2 ⎟ − bn ⎟⎟
⎝ ⎝ bn ⎠ ⎠
( )⎛
− s / 2 bn2 s / bn 2
⎛1 − e s / ( 2bn2 ) ⎞ b 2 ⎞
=e ⎜ n
b ( N − b ) + bn∫ e u/2
dW ±
( u ) + ⎜ ⎟ n⎟
⎝ ⎝ ⎠ ⎠
n
0

where W ( s ) is W ( s ) for s ≥ 0 and W ( s ) for s < 0. Note that uniformly for

± + −

s ≤ s0
− s / 2b2( ) ⎛ 1⎞
e = 1+ O ⎜ 2 ⎟.
⎝ bn ⎠
Further, since eu / 2 = 1 + O (1/ bn2 ) for u < s0 / bn2 ,
s / bn2 ⎛ ⎛ 1 ⎞⎞ ⎛ s ⎞
bn ∫ eu / 2 dW ± ( u ) = ⎜⎜1 + O ⎜ 2 ⎟ ⎟⎟ bnW ± ⎜ 2 ⎟ .
0
⎝ ⎝ bn ⎠ ⎠ ⎝ bn ⎠
Finally, for s ≤ s0 ,
⎛1 − e s / ( 2bn2 ) ⎞ b 2 = − s + O ⎛ 1 ⎞ .
⎜ ⎟ n ⎜ 2⎟
⎝ ⎠ 2 ⎝ bn ⎠
It follows that
⎛ ⎛ s ⎞ ⎞
bn ⎜⎜ X ⎜ 2 ⎟ − bn ⎟⎟
⎝ ⎝ bn ⎠ ⎠
⎛ ⎛ 1 ⎞⎞⎛ ⎛ s ⎞ s⎞ ⎛ 1 ⎞
= ⎜⎜ 1 + O ⎜ 2 ⎟ ⎟⎟ ⎜ bn ( N − bn ) + bnW ± ⎜ 2 ⎟ − ⎟ + O ⎜ 2 ⎟ .
⎜ ⎟
⎝ ⎝ bn ⎠ ⎠ ⎝ ⎝ bn ⎠ 2 ⎠ ⎝ bn ⎠
We write W ( s ) := bnW ± ( s / bn2 ) . Then W is also Brownian motion. We have
⎧⎪ n ⎛ ⎛ s ⎞ ⎞ ⎫⎪
⎨∨ bn ⎜⎜ X i ⎜ 2 ⎟ − bn ⎟⎟ ⎬
⎩⎪ i =1 ⎝ ⎝ bn ⎠ ⎠ ⎭⎪s∈
⎛ ⎛ 1 ⎞ ⎞ ⎪⎧ n ⎛ s ⎞ ⎪⎫ ⎛ 1⎞
= ⎜⎜ 1 + O ⎜ 2 ⎟ ⎟⎟ ⎨∨ ⎜ bn ( N i − bn ) + Wi ( s ) − ⎟ ⎬ + O ⎜ 2 ⎟ .
⎝ ⎝ bn ⎠ ⎠ ⎩⎪ i =1 ⎝ 2 ⎠ ⎪⎭ ⎝ bn ⎠

11 /14 11/21/2010
 (correction 2nd part made by MFH)
Hence the limit of (9.8.4.) is the same as that of
⎧n s⎫
(
⎨∨ bn ( N i − bn ) + Wi ( s ) − ⎬ .
2 ⎭s∈
) (9.8.5)
⎩ i =1

The rest of the proof runs as in the previous example.

One finds that the sequence of processes (9.8.5) converges weakly in
C [ − s0 , s0 ] hence in C ( ) , to

⎧∞ s⎫
(
⎨∨ log Z i + Wi ( s ) − ⎬
2 ⎭s∈
)
⎩ i =1
with {Zi}the point process from (9.8.1).
Note that the point process {Zi} and the random processes W i are
independent.
328 -7 independent of V independent of Y
329 1 u>0 x>0

Chapter 10: Estimation in C[0,1]

page line error/missing correction

332 -2 Theorem 10.2.1 Theorem 10.2.1 (de Haan and Lin (2003))
336 3 1 n 1 n
1 − F n , s ( x ) := ∑1
n j =1 { X i ( s )> x}
. 1 − F n , s ( x ) := ∑1
n j =1 { X j ( s )> x}
.

339 -3 Theorem 10.4.1 Theorem 10.4.1 (de Haan and Lin (2003))
341 -3 ζ n-k+1,n ζ n-k,n
352 6 υn (S) υ (S)
12 /14 11/21/2010
 (correction 2nd part made by MFH)

Appendix B: Regular Variation and Extensions

page line error correction

365 -1 and -3 (B.1.6) (B.1.16)
366 5 exp (integral} exp(integral)
370 -10 f ( t ) = exp [ log t ] f ( t ) = exp {− [ log t ]}

375 -10 (B.1.23) (B.2.12)

(B.1.24) (B.2.13)

-9 (B.1.24) (B.2.13)
376 -9 Hence f(t) is bounded for t ≥ t0 . Hence f(t) is locally bounded for t ≥ t0 .
379 3 f (∞ ) − f (t ) = f (∞) − f (t ) ∼

380 9 1 − xδ1 1 − x −δ1

(left side)
δ1 −δ1
381 -8 From Remark B.2.14(2) it From part 3 of the present proposition it follows
follows

Further and Updated References

S. Coles and J. Tawn: Modeling extreme multivariate events. J. Royal Statist. Soc. Ser. B 53, 277 – 392 (1991).
V. Dijk and L. de Haan: On the estimation of the exceedance probability of a high level. Order Statistics and non-parametrics: theory and applications. P.K.
Sen and I.A. Salama (Editors), 79-92. Elsevier, Amsterdam (1992)
L. de Haan and T. Lin: On convergence towards an extreme value distribution in C[0,1]. Ann. Prob. 29, 467-483 (2001).
L. de Haan and T. Lin: Weak consistency of extreme value estimators in C[0,1]. Ann. Statist. 31, 1996-2012 (2003).
L. de Haan and T.T. Pereira: Spatial Extremes: Models for the stationary case. Ann. Statist., 34, 146-168 (2006).
L. de Haan and H. Rootzén: On the estimation of high quantiles. J. Statist. Planning and Inference, 35, 1-13 (1993).
M. Schlather: Models for stationary max-stable random fields. Extremes, 5, 33-44 (2002).

13 /14 11/21/2010
 (correction 2nd part made by MFH)
View publication stats

R.L.Smith and I.Weissman: Maximum likelihood estimation of the lower tail of a probability distribution. J. Royal Statist. Soc. Ser. B 47, 285 – 298 (1985).

Data files

Last updated: November 21, 2010.

14 /14 11/21/2010